diff options
author | Thomas Martitz <kugel@rockbox.org> | 2012-01-07 19:56:09 +0100 |
---|---|---|
committer | Thomas Martitz <kugel@rockbox.org> | 2012-01-22 18:46:45 +0100 |
commit | a035261089403de259e74ce4dd196e2715138ed2 (patch) | |
tree | 440d75109a7ebfefef2a58076b38b1bfe9db8f63 /firmware/target/coldfire | |
parent | 8e8e978de6b6283b66a6829fa8c5e3b94674ce7d (diff) |
Move optimized memcpy and friends and strlen to firmware/asm,
using the new automatic-asm-picking infrastructure.
Diffstat (limited to 'firmware/target/coldfire')
-rw-r--r-- | firmware/target/coldfire/memcpy-coldfire.S | 682 | ||||
-rw-r--r-- | firmware/target/coldfire/memmove-coldfire.S | 670 | ||||
-rw-r--r-- | firmware/target/coldfire/memset-coldfire.S | 152 | ||||
-rw-r--r-- | firmware/target/coldfire/memset16-coldfire.S | 146 | ||||
-rw-r--r-- | firmware/target/coldfire/strlen-coldfire.S | 71 |
5 files changed, 0 insertions, 1721 deletions
diff --git a/firmware/target/coldfire/memcpy-coldfire.S b/firmware/target/coldfire/memcpy-coldfire.S deleted file mode 100644 index 9762e31e02..0000000000 --- a/firmware/target/coldfire/memcpy-coldfire.S +++ /dev/null @@ -1,682 +0,0 @@ -/*************************************************************************** - * __________ __ ___. - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ - * \/ \/ \/ \/ \/ - * $Id$ - * - * Copyright (C) 2004-2005 by Jens Arnold - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation; either version 2 - * of the License, or (at your option) any later version. - * - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY - * KIND, either express or implied. - * - ****************************************************************************/ -#include "config.h" - - .section .icode,"ax",@progbits - -#define FULLSPEED /* use burst writing for word aligned destinations */ - .align 2 - .global memcpy - .global __memcpy_fwd_entry - .type memcpy,@function - -/* Copies <length> bytes of data in memory from <source> to <dest> - * This version is optimized for speed - * - * arguments: - * (4,%sp) - destination address - * (8,%sp) - source address - * (12,%sp) - length - * - * return value: - * %d0 - destination address (like ANSI version) - * - * register usage: - * %a0 - current source address - * %a1 - current dest address - * %a2 - source end address (in line-copy loops) - * %d0 - data / scratch - * %d1 - source end address (byte and longword copy) / data / scratch - * %d2 - data / scratch - * %d3..%d7 - data - * - * For maximum speed this routine reads and writes whole lines using burst - * move (movem.l) where possible. For byte aligned destinations (long+1 and - * long+3) it writes longwords only. Same goes for word aligned destinations - * if FULLSPEED is undefined. - */ -memcpy: - move.l (4,%sp),%a1 /* Destination */ - move.l (8,%sp),%a0 /* Source */ - move.l (12,%sp),%d1 /* Length */ - -__memcpy_fwd_entry: - add.l %a0,%d1 /* %d1 = source end */ - - move.l %a0,%d0 - addq.l #7,%d0 - and.l #0xFFFFFFFC,%d0 /* %d0 = first source long bound + 4 */ - cmp.l %d0,%d1 /* at least one aligned longword to copy? */ - blo.w .bytes2_start /* no, jump directly to trailing byte loop */ - - subq.l #4,%d0 /* %d0 = first source long bound */ - cmp.l %a0,%d0 /* any bytes to copy? */ - jls .bytes1_end /* no: skip byte loop */ - - /* leading byte loop: copies 0..3 bytes */ -.bytes1_loop: - move.b (%a0)+,(%a1)+ /* copy byte */ - cmp.l %a0,%d0 /* runs %a0 up to first long bound */ - jhi .bytes1_loop - -.bytes1_end: - moveq.l #31,%d0 - add.l %a0,%d0 - and.l #0xFFFFFFF0,%d0 /* %d0 = first source line bound + 16 */ - cmp.l %d0,%d1 /* at least one aligned line to copy? */ - blo.w .long_start /* no: jump to longword copy loop */ - - lea.l (-28,%sp),%sp /* free up some registers */ - movem.l %d2-%d7/%a2,(%sp) - - moveq.l #16,%d2 - sub.l %d2,%d0 /* %d0 = first source line bound */ - move.l %d1,%a2 /* %a2 = end address */ - lea.l (-15,%a2),%a2 /* adjust end address for loops doing 16 bytes/ pass */ - move.l %a1,%d1 - moveq.l #3,%d2 /* mask */ - and.l %d2,%d1 - jmp.l (2,%pc,%d1.l*4) /* switch (dest_addr & 3) */ - bra.w .lines_do0_start - bra.w .lines_do1_start - bra.w .lines_do2_start - /* bra.w .lines_do3_start implicit */ - - /* byte aligned destination (long + 3): use line burst reads in main loop */ -.lines_do3_start: - moveq.l #24,%d1 /* shift count for shifting by 3 bytes */ - cmp.l %a0,%d0 /* any leading longwords? */ - jhi .lines_do3_head_start /* yes: leading longword copy */ - - movem.l (%a0),%d4-%d7 /* load first line */ - lea.l (16,%a0),%a0 - move.l %d4,%d2 - lsr.l %d1,%d2 /* get high byte of first longword */ - move.b %d2,(%a1)+ /* store byte */ - jra .lines_do3_entry /* jump into main loop */ - -.lines_do3_head_start: - move.l (%a0)+,%d7 /* load first longword */ - move.l %d7,%d2 - lsr.l %d1,%d2 /* get high byte */ - move.b %d2,(%a1)+ /* store byte */ - jra .lines_do3_head_entry /* jump into leading longword loop */ - -.lines_do3_head_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - move.l %d7,%d2 - lsr.l %d1,%d2 /* get high byte */ - or.l %d2,%d6 /* combine with old lower 3 bytes */ - move.l %d6,(%a1)+ /* store longword */ -.lines_do3_head_entry: - lsl.l #8,%d7 /* shift up lower 3 bytes */ - cmp.l %a0,%d0 /* runs %a0 up to first line bound */ - jhi .lines_do3_head_loop - -.lines_do3_loop: - move.l %d7,%d3 /* move last longword of old line away */ - movem.l (%a0),%d4-%d7 /* load new line */ - lea.l (16,%a0),%a0 - move.l %d4,%d2 - lsr.l %d1,%d2 /* get high byte of 1st longword */ - or.l %d2,%d3 /* combine with old lower 3 bytes */ - move.l %d3,(%a1)+ /* store longword */ -.lines_do3_entry: - lsl.l #8,%d4 /* shift up lower 3 bytes */ - move.l %d5,%d2 - lsr.l %d1,%d2 /* get high byte of 2nd longword */ - or.l %d2,%d4 /* combine with 1st lower 3 bytes */ - move.l %d4,(%a1)+ /* store longword */ - lsl.l #8,%d5 /* shift up lower 3 bytes */ - move.l %d6,%d2 - lsr.l %d1,%d2 /* get high byte of 3rd longword */ - or.l %d2,%d5 /* combine with 2nd lower 3 bytes */ - move.l %d5,(%a1)+ /* store longword */ - lsl.l #8,%d6 /* shift up lower 3 bytes */ - move.l %d7,%d2 - lsr.l %d1,%d2 /* get high byte of 4th longword */ - or.l %d2,%d6 /* combine with 3rd lower 3 bytes */ - move.l %d6,(%a1)+ /* store longword */ - lsl.l #8,%d7 /* shift up lower 3 bytes */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_do3_loop - - lea.l (12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a0,%a2 /* any trailing longwords? */ - jls .lines_do3_tail_end /* no: just store last lower 3 bytes */ - -.lines_do3_tail_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - move.l %d7,%d2 - lsr.l %d1,%d2 /* get high byte */ - or.l %d2,%d6 /* combine with old lower 3 bytes */ - move.l %d6,(%a1)+ /* store longword */ - lsl.l #8,%d7 /* shift up lower 3 bytes */ - cmp.l %a0,%a2 /* runs %a0 up to last long bound */ - jhi .lines_do3_tail_loop - -.lines_do3_tail_end: - swap %d7 /* get high word */ - move.w %d7,(%a1)+ /* store word */ - lsr.l %d1,%d7 /* get moved-up low byte */ - move.b %d7,(%a1)+ /* store byte */ - jra .lines_end - - /* byte aligned destination (long + 1): use line burst reads in main loop */ -.lines_do1_start: - moveq.l #24,%d1 /* shift count for shifting by 3 bytes */ - cmp.l %a0,%d0 /* any leading longwords? */ - jhi .lines_do1_head_start /* yes: leading longword copy */ - - movem.l (%a0),%d4-%d7 /* load first line */ - lea.l (16,%a0),%a0 - move.l %d4,%d2 /* first longword, bytes 3210 */ - lsr.l #8,%d2 /* first longword, bytes .321 */ - swap %d2 /* first longword, bytes 21.3 */ - move.b %d2,(%a1)+ /* store byte */ - swap %d2 /* first longword, bytes .321 */ - move.w %d2,(%a1)+ /* store word */ - jra .lines_do1_entry - -.lines_do1_head_start: - move.l (%a0)+,%d7 /* load first longword */ - move.l %d7,%d2 /* first longword, bytes 3210 */ - lsr.l #8,%d2 /* first longword, bytes .321 */ - swap %d2 /* first longword, bytes 21.3 */ - move.b %d2,(%a1)+ /* store byte */ - swap %d2 /* first longword, bytes .321 */ - move.w %d2,(%a1)+ /* store word */ - jra .lines_do1_head_entry - -.lines_do1_head_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - move.l %d7,%d2 - lsr.l #8,%d2 /* get upper 3 bytes */ - or.l %d2,%d6 /* combine with old low byte */ - move.l %d6,(%a1)+ /* store longword */ -.lines_do1_head_entry: - lsl.l %d1,%d7 /* shift up low byte */ - cmp.l %a0,%d0 /* runs %a0 up to first line bound */ - jhi .lines_do1_head_loop - -.lines_do1_loop: - move.l %d7,%d3 /* move last longword of old line away */ - movem.l (%a0),%d4-%d7 /* load new line */ - lea.l (16,%a0),%a0 - move.l %d4,%d2 - lsr.l #8,%d2 /* get upper 3 bytes of 1st longword */ - or.l %d2,%d3 /* combine with low byte of old longword */ - move.l %d3,(%a1)+ /* store longword */ -.lines_do1_entry: - lsl.l %d1,%d4 /* shift up low byte */ - move.l %d5,%d2 - lsr.l #8,%d2 /* get upper 3 bytes of 2nd longword */ - or.l %d2,%d4 /* combine with low byte of 1st longword */ - move.l %d4,(%a1)+ /* store longword */ - lsl.l %d1,%d5 /* shift up low byte */ - move.l %d6,%d2 - lsr.l #8,%d2 /* get upper 3 bytes of 3rd longword */ - or.l %d2,%d5 /* combine with low byte of 2nd longword */ - move.l %d5,(%a1)+ /* store longword */ - lsl.l %d1,%d6 /* shift up low byte */ - move.l %d7,%d2 - lsr.l #8,%d2 /* get upper 3 bytes of 4th longword */ - or.l %d2,%d6 /* combine with low byte of 4th longword */ - move.l %d6,(%a1)+ /* store longword */ - lsl.l %d1,%d7 /* shift up low byte */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_do1_loop - - lea.l (12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a0,%a2 /* any trailing longwords? */ - jls .lines_do1_tail_end /* no: just store last low byte */ - -.lines_do1_tail_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - move.l %d7,%d2 - lsr.l #8,%d2 /* get upper 3 bytes */ - or.l %d2,%d6 /* combine with old low byte */ - move.l %d6,(%a1)+ /* store longword */ - lsl.l %d1,%d7 /* shift up low byte */ - cmp.l %a0,%a2 /* runs %a0 up to last long bound */ - jhi .lines_do1_tail_loop - -.lines_do1_tail_end: - lsr.l %d1,%d7 /* get shifted-up low byte */ - move.b %d7,(%a1)+ /* store byte */ - jra .lines_end - - /* long aligned destination (line + 0/4/8/12): head */ -.lines_do0_head_loop: - move.l (%a0)+,(%a1)+ /* copy longword */ -.lines_do0_start: - cmp.l %a0,%d0 /* runs %a0 up to first line bound */ - jhi .lines_do0_head_loop - -.lines_do0_head_end: - move.l %a1,%d1 - lsr.l #2,%d1 - moveq.l #3,%d0 /* mask */ - and.l %d0,%d1 - moveq.l #16,%d0 /* address increment for one main loop pass */ - jmp.l (2,%pc,%d1.l*2) /* switch ((dest_addr >> 2) & 3) */ - bra.b .lines_lo0_start - bra.b .lines_lo4_start - bra.b .lines_lo8_start - /* bra.b .lines_lo12_start implicit */ - - /* long aligned destination (line + 12): use line bursts in the loop */ -.lines_lo12_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - move.l %d4,(%a1)+ /* store 1st longword */ - cmp.l %a0,%a2 /* any full lines? */ - jls .lines_lo12_end /* no: skip main loop */ - -.lines_lo12_loop: - move.l %d5,%d1 /* move last 3 longwords of old line away */ - move.l %d6,%d2 - move.l %d7,%d3 - movem.l (%a0),%d4-%d7 /* load new line */ - add.l %d0,%a0 - movem.l %d1-%d4,(%a1) /* store line (3 old + 1 new longwords) */ - add.l %d0,%a1 - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo12_loop - - /* long aligned destination (line + 0/4/8/12): tail */ -.lines_lo12_end: - move.l %d5,(%a1)+ /* store 3rd last longword */ -.lines_lo8_end: - move.l %d6,(%a1)+ /* store 2nd last longword */ -.lines_lo4_end: - move.l %d7,(%a1)+ /* store last longword */ -.lines_lo0_end: - lea.l (12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a0,%a2 /* any trailing longwords? */ - jls .lines_end /* no: get outta here */ - -.lines_do0_tail_loop: - move.l (%a0)+,(%a1)+ /* copy longword */ - cmp.l %a0,%a2 /* runs %a0 up to last long bound */ - jhi .lines_do0_tail_loop - - jra .lines_end - - /* line aligned destination: use line bursts in the loop */ -.lines_lo0_start: -.lines_lo0_loop: - movem.l (%a0),%d4-%d7 /* load line */ - add.l %d0,%a0 - movem.l %d4-%d7,(%a1) /* store line */ - add.l %d0,%a1 - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo0_loop - - jra .lines_lo0_end /* handle trailing longwords */ - - /* long aligned destination (line + 4): use line bursts in the loop */ -.lines_lo4_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - move.l %d4,(%a1)+ /* store 1st longword */ - move.l %d5,(%a1)+ /* store 2nd longword */ - move.l %d6,(%a1)+ /* store 3rd longword */ - cmp.l %a0,%a2 /* any full lines? */ - jls .lines_lo4_end /* no: skip main loop */ - -.lines_lo4_loop: - move.l %d7,%d3 /* move last longword of old line away */ - movem.l (%a0),%d4-%d7 /* load new line */ - add.l %d0,%a0 - movem.l %d3-%d6,(%a1) /* store line (1 old + 3 new longwords) */ - add.l %d0,%a1 - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo4_loop - - jra .lines_lo4_end /* handle trailing longwords */ - - /* long aligned destination (line + 8): use line bursts in the loop */ -.lines_lo8_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - move.l %d4,(%a1)+ /* store 1st longword */ - move.l %d5,(%a1)+ /* store 2nd longword */ - cmp.l %a0,%a2 - jls .lines_lo8_end - -.lines_lo8_loop: - move.l %d6,%d2 /* move last 2 longwords of old line away */ - move.l %d7,%d3 - movem.l (%a0),%d4-%d7 /* load new line */ - add.l %d0,%a0 - movem.l %d2-%d5,(%a1) /* store line (2 old + 2 new longwords) */ - add.l %d0,%a1 - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo8_loop - - jra .lines_lo8_end /* handle trailing longwords */ - -#ifdef FULLSPEED - - /* word aligned destination (line + 2/6/10/14): head */ -.lines_do2_start: - cmp.l %a0,%d0 /* any leading longwords? */ - jls .lines_do2_selector /* no: jump to mainloop selector */ - - move.l (%a0)+,%d7 /* load first longword */ - swap %d7 /* swap words */ - move.w %d7,(%a1)+ /* store high word */ - cmp.l %a0,%d0 /* any more longword? */ - jls .lines_do2_head_end /* no: skip head loop */ - -.lines_do2_head_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - swap %d7 /* swap words */ - move.w %d7,%d6 /* combine high word with old low word */ - move.l %d6,(%a1)+ /* store longword */ - cmp.l %a0,%d0 /* runs %a0 up to first line bound */ - jhi .lines_do2_head_loop - -.lines_do2_head_end: - swap %d7 /* undo swap */ - move.w %d7,(%a1)+ /* store word */ - -.lines_do2_selector: - move.l %a1,%d1 - lsr.l #2,%d1 - moveq.l #3,%d0 /* mask */ - and.l %d0,%d1 - moveq.l #16,%d0 /* address increment for one main loop pass */ - jmp.l (2,%pc,%d1.l*4) /* switch ((dest_addr >> 2) & 3) */ - bra.w .lines_lo2_start - bra.w .lines_lo6_start - bra.w .lines_lo10_start - /* bra.w .lines_lo14_start implicit */ - - /* word aligned destination (line + 14): use line bursts in the loop */ -.lines_lo14_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,(%a1)+ /* store word */ - jra .lines_lo14_entry /* jump into main loop */ - -.lines_lo14_loop: - move.l %d4,%d0 /* move old line away */ - move.l %d5,%d1 - move.l %d6,%d2 - move.l %d7,%d3 - movem.l (%a0),%d4-%d7 /* load new line */ - lea.l (16,%a0),%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,%d3 /* combine 1st high word with old low word */ - movem.l %d0-%d3,(%a1) /* store line */ - lea.l (16,%a1),%a1 -.lines_lo14_entry: - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo14_loop - - /* word aligned destination (line + 2/6/10/14): tail */ -.lines_lo14_end: - move.l %d4,(%a1)+ /* store third last longword */ -.lines_lo10_end: - move.l %d5,(%a1)+ /* store second last longword */ -.lines_lo6_end: - move.l %d6,(%a1)+ /* store last longword */ -.lines_lo2_end: - lea.l (12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a0,%a2 /* any trailing longwords? */ - jls .lines_do2_tail_end /* no: skip tail loop */ - -.lines_do2_tail_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - swap %d7 /* swap words */ - move.w %d7,%d6 /* combine high word with old low word */ - move.l %d6,(%a1)+ /* store longword */ - cmp.l %a0,%a2 /* runs %a0 up to last long bound */ - jhi .lines_do2_tail_loop - -.lines_do2_tail_end: - swap %d7 /* undo swap */ - move.w %d7,(%a1)+ /* store last word */ - jra .lines_end - - /* word aligned destination (line + 2): use line bursts in the loop */ -.lines_lo2_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,(%a1)+ /* store high word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3nd high word with 2nd low word */ - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - move.l %d4,(%a1)+ /* store 1st longword */ - move.l %d5,(%a1)+ /* store 2nd longword */ - move.l %d6,(%a1)+ /* store 3rd longword */ - cmp.l %a0,%a2 /* any full lines? */ - jls .lines_lo2_end /* no: skip main loop */ - -.lines_lo2_loop: - move.l %d7,%d3 /* move last longword of old line away */ - movem.l (%a0),%d4-%d7 /* load line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,%d3 /* combine 1st high word with old low word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - movem.l %d3-%d6,(%a1) /* store line */ - add.l %d0,%a1 - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo2_loop - - jra .lines_lo2_end /* handle trailing longwords */ - - /* word aligned destination (line + 6): use line bursts in the loop */ -.lines_lo6_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,(%a1)+ /* store high word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - move.l %d4,(%a1)+ /* store 1st longword */ - move.l %d5,(%a1)+ /* store 2nd longword */ - jra .lines_lo6_entry /* jump into main loop */ - -.lines_lo6_loop: - move.l %d6,%d2 /* move last 2 longwords of old line away */ - move.l %d7,%d3 - movem.l (%a0),%d4-%d7 /* load line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,%d3 /* combine 1st high word with old low word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - movem.l %d2-%d5,(%a1) /* store line */ - add.l %d0,%a1 -.lines_lo6_entry: - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo6_loop - - jra .lines_lo6_end /* handle trailing longwords */ - - /* word aligned destination (line + 10): use line bursts in the loop */ -.lines_lo10_start: - movem.l (%a0),%d4-%d7 /* load first line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,(%a1)+ /* store high word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - move.l %d4,(%a1)+ /* store 1st longword */ - jra .lines_lo10_entry /* jump into main loop */ - -.lines_lo10_loop: - move.l %d5,%d1 /* move last 3 longwords of old line away */ - move.l %d6,%d2 - move.l %d7,%d3 - movem.l (%a0),%d4-%d7 /* load line */ - add.l %d0,%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,%d3 /* combine 1st high word with old low word */ - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - movem.l %d1-%d4,(%a1) /* store line */ - add.l %d0,%a1 -.lines_lo10_entry: - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_lo10_loop - - jra .lines_lo10_end /* handle trailing longwords */ - -#else /* !FULLSPEED */ - - /* word aligned destination (long + 2): use line burst reads in the loop */ -.lines_do2_start: - cmp.l %a0,%d0 /* any leading longwords? */ - jhi .lines_do2_head_start /* yes: leading longword copy */ - - movem.l (%a0),%d4-%d7 /* load first line */ - lea.l (16,%a0),%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,(%a1)+ /* store high word */ - jra .lines_do2_entry /* jump into main loop */ - -.lines_do2_head_start: - move.l (%a0)+,%d7 /* load first longword */ - swap %d7 /* swap words */ - move.w %d7,(%a1)+ /* store high word */ - cmp.l %a0,%d0 /* any full longword? */ - jls .lines_do2_loop /* no: skip head loop */ - -.lines_do2_head_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - swap %d7 /* swap words */ - move.w %d7,%d6 /* combine high word with old low word */ - move.l %d6,(%a1)+ /* store longword */ - cmp.l %a0,%d0 /* runs %a0 up to first line bound */ - jhi .lines_do2_head_loop - -.lines_do2_loop: - move.l %d7,%d3 /* move last longword of old line away */ - movem.l (%a0),%d4-%d7 /* load line */ - lea.l (16,%a0),%a0 - swap %d4 /* swap words of 1st long */ - move.w %d4,%d3 /* combine 1st high word with old low word */ - move.l %d3,(%a1)+ /* store 1st longword */ -.lines_do2_entry: - swap %d5 /* swap words of 2nd long */ - move.w %d5,%d4 /* combine 2nd high word with 1st low word */ - move.l %d4,(%a1)+ /* store 2nd longword */ - swap %d6 /* swap words of 3rd long */ - move.w %d6,%d5 /* combine 3rd high word with 2nd low word */ - move.l %d5,(%a1)+ /* store 3rd longword */ - swap %d7 /* swap words of 4th long */ - move.w %d7,%d6 /* combine 4th high word with 3rd low word */ - move.l %d6,(%a1)+ /* store 4th longword */ - cmp.l %a0,%a2 /* runs %a0 up to last line bound */ - jhi .lines_do2_loop - -.lines_do2_end: - lea.l (12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a0,%a2 /* any trailing longwords? */ - jls .lines_do2_tail_end /* no: skip tail loop */ - -.lines_do2_tail_loop: - move.l %d7,%d6 /* move old longword away */ - move.l (%a0)+,%d7 /* load new longword */ - swap %d7 /* swap words */ - move.w %d7,%d6 /* combine high word with old low word */ - move.l %d6,(%a1)+ /* store longword */ - cmp.l %a0,%a2 /* runs %a0 up to last long bound */ - jhi .lines_do2_tail_loop - -.lines_do2_tail_end: - swap %d7 /* undo swap */ - move.w %d7,(%a1)+ /* store last word */ - /* jra .lines_end implicit */ - -#endif /* !FULLSPEED */ - -.lines_end: - addq.l #3,%a2 /* readjust end address */ - move.l %a2,%d1 /* end address in %d1 again */ - movem.l (%sp),%d2-%d7/%a2 /* restore registers */ - lea.l (28,%sp),%sp - jra .bytes2_start /* jump to trailing byte loop */ - -.long_start: - subq.l #3,%d1 /* adjust end address for doing 4 bytes/ pass */ - - /* longword copy loop - no lines */ -.long_loop: - move.l (%a0)+,(%a1)+ /* copy longword (write can be unaligned) */ - cmp.l %a0,%d1 /* runs %a0 up to last long bound */ - jhi .long_loop - - addq.l #3,%d1 /* readjust end address */ - cmp.l %a0,%d1 /* any bytes left? */ - jls .bytes2_end /* no: skip trailing byte loop */ - - /* trailing byte loop */ -.bytes2_loop: - move.b (%a0)+,(%a1)+ /* copy byte */ -.bytes2_start: - cmp.l %a0,%d1 /* runs %a0 up to end address */ - jhi .bytes2_loop - -.bytes2_end: - move.l (4,%sp),%d0 /* return destination */ - rts - -.end: - .size memcpy,.end-memcpy diff --git a/firmware/target/coldfire/memmove-coldfire.S b/firmware/target/coldfire/memmove-coldfire.S deleted file mode 100644 index 736cd619e1..0000000000 --- a/firmware/target/coldfire/memmove-coldfire.S +++ /dev/null @@ -1,670 +0,0 @@ -/*************************************************************************** - * __________ __ ___. - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ - * \/ \/ \/ \/ \/ - * $Id$ - * - * Copyright (C) 2006 by Jens Arnold - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation; either version 2 - * of the License, or (at your option) any later version. - * - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY - * KIND, either express or implied. - * - ****************************************************************************/ -#include "config.h" - - .section .icode,"ax",@progbits - -#define FULLSPEED /* use burst writing for word aligned destinations */ - .align 2 - .global memmove - .type memmove,@function - -/* Moves <length> bytes of data in memory from <source> to <dest> - * Regions may overlap. - * This version is optimized for speed, and needs the corresponding memcpy - * implementation for the forward copy branch. - * - * arguments: - * (4,%sp) - destination address - * (8,%sp) - source address - * (12,%sp) - length - * - * return value: - * %d0 - destination address (like ANSI version) - * - * register usage: - * %a0 - current source address - * %a1 - current dest address - * %a2 - source start address (in line-copy loops) - * %d0 - source start address (byte and longword copy) / data / scratch - * %d1 - data / scratch - * %d2 - data / scratch - * %d3..%d7 - data - * - * For maximum speed this routine reads and writes whole lines using burst - * move (movem.l) where possible. For byte aligned destinations (long-1 and - * long-3) it writes longwords only. Same goes for word aligned destinations - * if FULLSPEED is undefined. - */ -memmove: - move.l (4,%sp),%a1 /* Destination */ - move.l (8,%sp),%a0 /* Source */ - move.l (12,%sp),%d1 /* Length */ - - cmp.l %a0,%a1 - bhi.b .backward /* dest > src -> backward copy */ - jmp __memcpy_fwd_entry - -.backward: - move.l %a0,%d0 /* %d0 = source start */ - add.l %d1,%a0 /* %a0 = source end */ - add.l %d1,%a1 /* %a1 = destination end */ - - move.l %a0,%d1 - and.l #0xFFFFFFFC,%d1 /* %d1 = last source long bound */ - subq.l #4,%d1 - cmp.l %d0,%d1 /* at least one aligned longword to copy? */ - blo.w .bytes2r_start - - addq.l #4,%d1 /* %d1 = last source long bound */ - cmp.l %d1,%a0 /* any bytes to copy */ - jls .bytes1r_end /* no: skip byte loop */ - - /* leading byte loop: copies 0..3 bytes */ -.bytes1r_loop: - move.b -(%a0),-(%a1) /* copy byte */ - cmp.l %d1,%a0 /* runs %a0 down to last long bound */ - jhi .bytes1r_loop - -.bytes1r_end: - moveq.l #-16,%d1 - add.l %a0,%d1 - and.l #0xFFFFFFF0,%d1 /* %d1 = last source line bound - 16 */ - cmp.l %d0,%d1 /* at least one aligned line to copy? */ - blo.w .longr_start /* no: jump to longword copy loop */ - - lea.l (-28,%sp),%sp /* free up some registers */ - movem.l %d2-%d7/%a2,(%sp) - - moveq.l #16,%d2 - add.l %d2,%d1 /* %d1 = last source line bound */ - move.l %d0,%a2 /* %a2 = start address */ - lea.l (15,%a2),%a2 /* adjust start address for loops doing 16 bytes/pass */ - move.l %a1,%d0 - moveq.l #3,%d2 /* mask */ - and.l %d2,%d0 - jmp.l (2,%pc,%d0.l*4) /* switch (dest_addr & 3) */ - bra.w .lines_do0r_start - bra.w .lines_do1r_start - bra.w .lines_do2r_start - /* bra.w .lines_do3r_start implicit */ - - /* byte aligned destination (long - 1): use line burst reads in main loop */ -.lines_do3r_start: - moveq.l #24,%d0 /* shift count for shifting by 3 bytes */ - cmp.l %d1,%a0 /* any leading longwords? */ - jhi .lines_do3r_head_start /* yes: leading longword copy */ - - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load initial line */ - move.l %d6,%d2 /* last longword, bytes 3210 */ - move.b %d2,-(%a1) /* store byte */ - lsr.l #8,%d2 /* last longword, bytes .321 */ - move.w %d2,-(%a1) /* store word */ - jra .lines_do3r_entry - -.lines_do3r_head_start: - move.l -(%a0),%d3 /* load initial longword */ - move.l %d3,%d2 /* bytes 3210 */ - move.b %d2,-(%a1) /* store byte */ - lsr.l #8,%d2 /* bytes .321 */ - move.w %d2,-(%a1) /* store word */ - jra .lines_do3r_head_entry - -.lines_do3r_head_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.l %d3,%d2 - lsl.l #8,%d2 /* get bytes 210. */ - or.l %d2,%d4 /* combine with old high byte */ - move.l %d4,-(%a1) /* store longword */ -.lines_do3r_head_entry: - lsr.l %d0,%d3 /* shift down high byte */ - cmp.l %d1,%a0 /* run %a0 down to last line bound */ - jhi .lines_do3r_head_loop - -.lines_do3r_loop: - move.l %d3,%d7 /* move first longword of last line away */ - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load new line */ - move.l %d6,%d2 - lsl.l #8,%d2 /* get bytes 210. of 4th longword */ - or.l %d2,%d7 /* combine with high byte of old longword */ - move.l %d7,-(%a1) /* store longword */ -.lines_do3r_entry: - lsr.l %d0,%d6 /* shift down high byte */ - move.l %d5,%d2 - lsl.l #8,%d2 /* get bytes 210. of 3rd longword */ - or.l %d2,%d6 /* combine with high byte of 4th longword */ - move.l %d6,-(%a1) /* store longword */ - lsr.l %d0,%d5 /* shift down high byte */ - move.l %d4,%d2 - lsl.l #8,%d2 /* get bytes 210. of 2nd longword */ - or.l %d2,%d5 /* combine with high byte or 3rd longword */ - move.l %d5,-(%a1) /* store longword */ - lsr.l %d0,%d4 /* shift down high byte */ - move.l %d3,%d2 - lsl.l #8,%d2 /* get bytes 210. of 1st longword */ - or.l %d2,%d4 /* combine with high byte of 2nd longword */ - move.l %d4,-(%a1) /* store longword */ - lsr.l %d0,%d3 /* shift down high byte */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_do3r_loop - - lea.l (-12,%a2),%a2 /* readjust start address for doing longwords */ - cmp.l %a2,%a0 /* any trailing longwords? */ - jls .lines_do3r_tail_end /* no: just store last high byte */ - -.lines_do3r_tail_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.l %d3,%d2 - lsl.l #8,%d2 /* get bytes 210. */ - or.l %d2,%d4 /* combine with old high byte */ - move.l %d4,-(%a1) /* store longword */ - lsr.l %d0,%d3 /* shift down high byte */ - cmp.l %a2,%a0 /* run %a0 down to first long bound */ - jhi .lines_do3r_tail_loop - -.lines_do3r_tail_end: - move.b %d3,-(%a1) /* store shifted-down high byte */ - jra .linesr_end - - /* byte aligned destination (long - 3): use line burst reads in main loop */ -.lines_do1r_start: - moveq.l #24,%d0 /* shift count for shifting by 3 bytes */ - cmp.l %d1,%a0 /* any leading longwords? */ - jhi .lines_do1r_head_start /* yes: leading longword copy */ - - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load initial line */ - move.b %d6,-(%a1) /* store low byte of last longword */ - jra .lines_do1r_entry - -.lines_do1r_head_start: - move.l -(%a0),%d3 /* load initial longword */ - move.b %d3,-(%a1) /* store low byte */ - jra .lines_do1r_head_entry - -.lines_do1r_head_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.l %d3,%d2 - lsl.l %d0,%d2 /* get low byte */ - or.l %d2,%d4 /* combine with old bytes .321 */ - move.l %d4,-(%a1) /* store longword */ -.lines_do1r_head_entry: - lsr.l #8,%d3 /* get bytes .321 */ - cmp.l %d1,%a0 /* run %a0 down to last line bound */ - jhi .lines_do1r_head_loop - -.lines_do1r_loop: - move.l %d3,%d7 /* move first longword of old line away */ - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load new line */ - move.l %d6,%d2 - lsl.l %d0,%d2 /* get low byte of 4th longword */ - or.l %d2,%d7 /* combine with bytes .321 of old longword */ - move.l %d7,-(%a1) /* store longword */ -.lines_do1r_entry: - lsr.l #8,%d6 /* get bytes .321 */ - move.l %d5,%d2 - lsl.l %d0,%d2 /* get low byte of 3rd longword */ - or.l %d2,%d6 /* combine with bytes .321 of 4th longword */ - move.l %d6,-(%a1) /* store longword */ - lsr.l #8,%d5 /* get bytes .321 */ - move.l %d4,%d2 - lsl.l %d0,%d2 /* get low byte of 2nd longword */ - or.l %d2,%d5 /* combine with bytes .321 of 3rd longword */ - move.l %d5,-(%a1) /* store longword */ - lsr.l #8,%d4 /* get bytes .321 */ - move.l %d3,%d2 - lsl.l %d0,%d2 /* get low byte of 1st longword */ - or.l %d2,%d4 /* combine with bytes .321 of 2nd longword */ - move.l %d4,-(%a1) /* store longword */ - lsr.l #8,%d3 /* get bytes .321 */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_do1r_loop - - lea.l (-12,%a2),%a2 /* readjust start address for doing longwords */ - cmp.l %a2,%a0 /* any trailing longwords? */ - jls .lines_do1r_tail_end /* no: just store last high byte */ - -.lines_do1r_tail_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.l %d3,%d2 - lsl.l %d0,%d2 /* get low byte */ - or.l %d2,%d4 /* combine with old bytes .321 */ - move.l %d4,-(%a1) /* store longword */ - lsr.l #8,%d3 /* get bytes .321 */ - cmp.l %a2,%a0 /* run %a0 down to first long bound */ - jhi .lines_do1r_tail_loop - -.lines_do1r_tail_end: - move.w %d3,-(%a1) /* store word 21 */ - swap %d3 - move.b %d3,-(%a1) /* store byte 3 */ - jra .linesr_end - - /* long aligned destination (line - 0/4/8/12): head */ -.lines_do0r_head_loop: - move.l -(%a0),-(%a1) /* copy longword */ -.lines_do0r_start: - cmp.l %d1,%a0 /* run %a0 down to last line bound */ - jhi .lines_do0r_head_loop - -.lines_do0r_head_end: - move.l %a1,%d1 - lsr.l #2,%d1 - moveq.l #3,%d0 /* mask */ - and.l %d0,%d1 - moveq.l #16,%d0 /* address decrement for one main loop pass */ - jmp.l (2,%pc,%d1.l*2) /* switch ((dest_addr >> 2) & 3) */ - bra.b .lines_lo0r_start - bra.b .lines_lo4r_start - bra.b .lines_lo8r_start - /* bra.b .lines_lo12r_start implicit */ - - /* long aligned destination (line - 4): use line bursts in the loop */ -.lines_lo12r_start: - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load initial line */ - move.l %d4,-(%a1) /* store 4th longword */ - move.l %d3,-(%a1) /* store 3rd longword */ - move.l %d2,-(%a1) /* store 2nd longword */ - cmp.l %a2,%a0 /* any full lines? */ - jls .lines_lo12r_end /* no: skip main loop */ - -.lines_lo12r_loop: - move.l %d1,%d5 /* move first longword of old line away */ - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load new line */ - sub.l %d0,%a1 - movem.l %d2-%d5,(%a1) /* store line (1 old + 3 new longwords */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo12r_loop - - jra .lines_lo12r_end /* handle trailing longwords */ - - /* line aligned destination: use line bursts in the loop */ -.lines_lo0r_start: -.lines_lo0r_loop: - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load line */ - sub.l %d0,%a1 - movem.l %d1-%d4,(%a1) /* store line */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo0r_loop - - jra .lines_lo0r_end /* handle trailing longwords */ - - /* long aligned destination (line - 8): use line bursts in the loop */ -.lines_lo8r_start: - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load initial line */ - move.l %d4,-(%a1) /* store 4th longword */ - move.l %d3,-(%a1) /* store 3rd longword */ - cmp.l %a2,%a0 /* any full lines? */ - jls .lines_lo8r_end /* no: skip main loop */ - -.lines_lo8r_loop: - move.l %d2,%d6 /* move first 2 longwords of old line away */ - move.l %d1,%d5 - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load new line */ - sub.l %d0,%a1 - movem.l %d3-%d6,(%a1) /* store line (2 old + 2 new longwords */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo8r_loop - - jra .lines_lo8r_end /* handle trailing longwords */ - - /* long aligned destination (line - 12): use line bursts in the loop */ -.lines_lo4r_start: - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load initial line */ - move.l %d4,-(%a1) /* store 4th longword */ - cmp.l %a2,%a0 /* any full lines? */ - jls .lines_lo4r_end /* no: skip main loop */ - -.lines_lo4r_loop: - move.l %d3,%d7 /* move first 3 longwords of old line away */ - move.l %d2,%d6 - move.l %d1,%d5 - sub.l %d0,%a0 - movem.l (%a0),%d1-%d4 /* load new line */ - sub.l %d0,%a1 - movem.l %d4-%d7,(%a1) /* store line (3 old + 1 new longwords */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo4r_loop - - /* long aligned destination (line - 0/4/8/12): tail */ -.lines_lo4r_end: - move.l %d3,-(%a1) /* store 3rd last longword */ -.lines_lo8r_end: - move.l %d2,-(%a1) /* store 2nd last longword */ -.lines_lo12r_end: - move.l %d1,-(%a1) /* store last longword */ -.lines_lo0r_end: - lea.l (-12,%a2),%a2 /* readjust end address for doing longwords */ - cmp.l %a2,%a0 /* any trailing longwords? */ - jls .linesr_end /* no: get outta here */ - -.lines_do0r_tail_loop: - move.l -(%a0),-(%a1) /* copy longword */ - cmp.l %a2,%a0 /* run %a0 down to first long bound */ - jhi .lines_do0r_tail_loop - - jra .linesr_end - -#ifdef FULLSPEED - /* word aligned destination (line - 2/6/10/14): head */ -.lines_do2r_start: - cmp.l %d1,%a0 /* any leading longwords? */ - jls .lines_do2r_selector /* no: jump to mainloop selector */ - - move.l -(%a0),%d3 /* load initial longword */ - move.w %d3,-(%a1) /* store low word */ - cmp.l %d1,%a0 /* any more longwords? */ - jls .lines_do2r_head_end /* no: skip head loop */ - -.lines_do2r_head_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.w %d3,%d4 /* combine low word with old high word */ - swap %d4 /* swap words */ - move.l %d4,-(%a1) /* store longword */ - cmp.l %d1,%a0 /* run %a0 down to last line bound */ - jhi .lines_do2r_head_loop - -.lines_do2r_head_end: - swap %d3 /* get high word */ - move.w %d3,-(%a1) /* and store it */ - -.lines_do2r_selector: - move.l %a1,%d1 - lsr.l #2,%d1 - moveq.l #3,%d0 /* mask */ - and.l %d0,%d1 - moveq.l #16,%d7 /* address decrement for one main loop pass */ - jmp.l (2,%pc,%d1.l*4) /* switch ((dest_addr >> 2) & 3) */ - bra.w .lines_lo2r_start - bra.w .lines_lo6r_start - bra.w .lines_lo10r_start - /* bra.w .lines_lo14r_start implicit */ - - /* word aligned destination (line - 2): use line bursts in the loop */ -.lines_lo14r_start: - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load initial line */ - move.w %d3,-(%a1) /* store last low word */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - move.w %d0,%d1 /* combine 1st low word with 2nd high word */ - swap %d1 /* swap words of 1st long */ - move.l %d3,-(%a1) /* store 3rd longword */ - move.l %d2,-(%a1) /* store 2nd longword */ - move.l %d1,-(%a1) /* store 1st longword */ - cmp.l %a2,%a0 /* any full lines? */ - jls .lines_lo14r_end /* no: skip main loop */ - -.lines_lo14r_loop: - move.l %d0,%d4 /* move first longword of old line away */ - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load line */ - move.w %d3,%d4 /* combine 4th low word with old high word */ - swap %d4 /* swap words of 4th long */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - move.w %d0,%d1 /* combine 1st low word with 2nd high word */ - swap %d1 /* swap words of 1st long */ - sub.l %d7,%a1 - movem.l %d1-%d4,(%a1) /* store line */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo14r_loop - - jra .lines_lo14r_end /* handle trailing longwords */ - - /* word aligned destination (line - 6): use line bursts in the loop */ -.lines_lo10r_start: - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load initial line */ - move.w %d3,-(%a1) /* store last low word */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - move.l %d3,-(%a1) /* store 3rd longword */ - move.l %d2,-(%a1) /* store 2nd longword */ - jra .lines_lo10r_entry /* jump into main loop */ - -.lines_lo10r_loop: - move.l %d0,%d4 /* move first 2 longwords of old line away */ - move.l %d1,%d5 - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load line */ - move.w %d3,%d4 /* combine 4th low word with old high word */ - swap %d4 /* swap words of 4th long */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - sub.l %d7,%a1 - movem.l %d2-%d5,(%a1) /* store line */ -.lines_lo10r_entry: - move.w %d0,%d1 /* combine 1st low word with 2nd high word */ - swap %d1 /* swap words of 1st long */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo10r_loop - - jra .lines_lo10r_end /* handle trailing longwords */ - - /* word aligned destination (line - 10): use line bursts in the loop */ -.lines_lo6r_start: - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load initial line */ - move.w %d3,-(%a1) /* store last low word */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.l %d3,-(%a1) /* store 3rd longword */ - jra .lines_lo6r_entry /* jump into main loop */ - -.lines_lo6r_loop: - move.l %d0,%d4 /* move first 3 longwords of old line away */ - move.l %d1,%d5 - move.l %d2,%d6 - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load line */ - move.w %d3,%d4 /* combine 4th low word with old high word */ - swap %d4 /* swap words of 4th long */ - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - sub.l %d7,%a1 - movem.l %d3-%d6,(%a1) /* store line */ -.lines_lo6r_entry: - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - move.w %d0,%d1 /* combine 1st low word with 2nd high word */ - swap %d1 /* swap words of 1st long */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo6r_loop - - jra .lines_lo6r_end /* handle trailing longwords */ - - /* word aligned destination (line - 14): use line bursts in the loop */ -.lines_lo2r_start: - sub.l %d7,%a0 - movem.l (%a0),%d0-%d3 /* load initial line */ - move.w %d3,-(%a1) /* store last low word */ - jra .lines_lo2r_entry /* jump into main loop */ - -.lines_lo2r_loop: - move.l %d0,%d4 /* move old line away */ - move.l %d1,%d5 - move.l %d2,%d6 - move.l %d3,%d7 - lea.l (-16,%a0),%a0 - movem.l (%a0),%d0-%d3 /* load line */ - move.w %d3,%d4 /* combine 4th low word with old high word */ - swap %d4 /* swap words of 4th long */ - lea.l (-16,%a1),%a1 - movem.l %d4-%d7,(%a1) /* store line */ -.lines_lo2r_entry: - move.w %d2,%d3 /* combine 3rd low word with 4th high word */ - swap %d3 /* swap words of 3rd long */ - move.w %d1,%d2 /* combine 2nd low word with 3rd high word */ - swap %d2 /* swap words of 2nd long */ - move.w %d0,%d1 /* combine 1st low word with 2nd high word */ - swap %d1 /* swap words of 1st long */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_lo2r_loop - - /* word aligned destination (line - 2/6/10/14): tail */ -.lines_lo2r_end: - move.l %d3,-(%a1) /* store third last longword */ -.lines_lo6r_end: - move.l %d2,-(%a1) /* store second last longword */ -.lines_lo10r_end: - move.l %d1,-(%a1) /* store last longword */ -.lines_lo14r_end: - lea.l (-12,%a2),%a2 /* readjust start address for doing longwords */ - cmp.l %a2,%a0 /* any trailing longwords? */ - jls .lines_do2r_tail_end /* no: skip tail loop */ - -.lines_do2r_tail_loop: - move.l %d0,%d1 /* move old longword away */ - move.l -(%a0),%d0 /* load new longword */ - move.w %d0,%d1 /* combine low word with old high word */ - swap %d1 /* swap words */ - move.l %d1,-(%a1) /* store longword */ - cmp.l %a2,%a0 /* run %a0 down to first long bound */ - jhi .lines_do2r_tail_loop - -.lines_do2r_tail_end: - swap %d0 /* get final high word */ - move.w %d0,-(%a1) /* store it */ - /* jra .linesr_end implicit */ - -#else /* !FULLSPEED */ - - /* word aligned destination (long - 2): use line burst reads in the loop */ -.lines_do2r_start: - cmp.l %d1,%a0 /* any leading longwords? */ - jhi .lines_do2r_head_start /* yes: leading longword copy */ - - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load initial line */ - move.w %d6,-(%a1) /* store last low word */ - jra .lines_do2r_entry /* jump into main loop */ - -.lines_do2r_head_start: - move.l -(%a0),%d3 /* load initial longword */ - move.w %d3,-(%a1) /* store low word */ - cmp.l %d1,%a0 /* any full longword? */ - jls .lines_do2r_loop /* no: skip head loop */ - -.lines_do2r_head_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.w %d3,%d4 /* combine low word with old high word */ - swap %d4 /* swap words */ - move.l %d4,-(%a1) /* store longword */ - cmp.l %d1,%a0 /* run %a0 down to last line bound */ - jhi .lines_do2r_head_loop - -.lines_do2r_loop: - move.l %d3,%d7 /* move first longword of old line away */ - lea.l (-16,%a0),%a0 - movem.l (%a0),%d3-%d6 /* load line */ - move.w %d6,%d7 /* combine 4th low word with old high word */ - swap %d7 /* swap words of 4th long */ - move.l %d7,-(%a1) /* store 4th longword */ -.lines_do2r_entry: - move.w %d5,%d6 /* combine 3rd low word with 4th high word */ - swap %d6 /* swap words of 3rd long */ - move.l %d6,-(%a1) /* store 3rd longword */ - move.w %d4,%d5 /* combine 2nd low word with 3rd high word */ - swap %d5 /* swap words of 2nd long */ - move.l %d5,-(%a1) /* store 2nd longword */ - move.w %d3,%d4 /* combine 1st low word with 2nd high word */ - swap %d4 /* swap words of 1st long */ - move.l %d4,-(%a1) /* store 1st longword */ - cmp.l %a2,%a0 /* run %a0 down to first line bound */ - jhi .lines_do2r_loop - -.lines_do2r_end: - lea.l (-12,%a2),%a2 /* readjust start address for doing longwords */ - cmp.l %a2,%a0 /* any trailing longwords? */ - jls .lines_do2r_tail_end /* no: skip tail loop */ - -.lines_do2r_tail_loop: - move.l %d3,%d4 /* move old longword away */ - move.l -(%a0),%d3 /* load new longword */ - move.w %d3,%d4 /* combine low word with old high word */ - swap %d4 /* swap words */ - move.l %d4,-(%a1) /* store longword */ - cmp.l %a2,%a0 /* run %a0 down to first long bound */ - jhi .lines_do2r_tail_loop - -.lines_do2r_tail_end: - swap %d3 /* get final high word */ - move.w %d3,-(%a1) /* store it */ - /* jra .linesr_end implicit */ - -#endif /* !FULLSPEED */ - -.linesr_end: - subq.l #3,%a2 /* readjust end address */ - move.l %a2,%d0 /* start address in %d0 again */ - movem.l (%sp),%d2-%d7/%a2 /* restore registers */ - lea.l (28,%sp),%sp - jra .bytes2r_start /* jump to trailing byte loop */ - -.longr_start: - addq.l #3,%d0 /* adjust start address for doing 4 bytes/ pass */ - - /* longword copy loop - no lines */ -.longr_loop: - move.l -(%a0),-(%a1) /* copy longword (write can be unaligned) */ - cmp.l %d0,%a0 /* runs %a0 down to first long bound */ - jhi .longr_loop - - subq.l #3,%d0 /* readjust start address */ - cmp.l %d0,%a0 /* any bytes left? */ - jls .bytes2r_end /* no: skip trailing byte loop */ - - /* trailing byte loop */ -.bytes2r_loop: - move.b -(%a0),-(%a1) /* copy byte */ -.bytes2r_start: - cmp.l %d0,%a0 /* runs %a0 down to start address */ - jhi .bytes2r_loop - -.bytes2r_end: - rts /* returns start address */ - -.end: - .size memmove,.end-memmove diff --git a/firmware/target/coldfire/memset-coldfire.S b/firmware/target/coldfire/memset-coldfire.S deleted file mode 100644 index 839b305a05..0000000000 --- a/firmware/target/coldfire/memset-coldfire.S +++ /dev/null @@ -1,152 +0,0 @@ -/*************************************************************************** - * __________ __ ___. - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ - * \/ \/ \/ \/ \/ - * $Id$ - * - * Copyright (C) 2004 by Jens Arnold - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation; either version 2 - * of the License, or (at your option) any later version. - * - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY - * KIND, either express or implied. - * - ****************************************************************************/ -#include "config.h" - - .section .icode,"ax",@progbits - - .align 2 - .global memset - .type memset,@function - -/* Fills a memory region with specified byte value - * This version is optimized for speed - * - * arguments: - * (4,%sp) - start address - * (8,%sp) - data - * (12,%sp) - length - * - * return value: - * %d0 - start address (like ANSI version) - * - * register usage: - * %d0 - data (spread to all 4 bytes when using long stores) - * %d1 - temporary / data (for burst transfer) - * %d2 - data (for burst transfer) - * %d3 - data (for burst transfer) - * %a0 - start address - * %a1 - current address (runs down from end to start) - * - * For maximum speed this routine uses both long stores and burst mode, - * storing whole lines with movem.l. The routine fills memory from end - * to start in order to ease returning the start address. - */ -memset: - move.l (4,%sp),%a0 /* start address */ - move.l (8,%sp),%d0 /* data */ - move.l (12,%sp),%a1 /* length */ - add.l %a0,%a1 /* %a1 = end address */ - - move.l %a0,%d1 - addq.l #7,%d1 - and.l #0xFFFFFFFC,%d1 /* %d1 = first long bound + 4 */ - cmp.l %d1,%a1 /* at least one aligned longword to fill? */ - blo.b .no_longs /* no, jump directly to byte loop */ - - and.l #0xFF,%d0 /* start: spread data to all 4 bytes */ - move.l %d0,%d1 - lsl.l #8,%d1 - or.l %d1,%d0 /* data now in 2 lower bytes of %d0 */ - move.l %d0,%d1 - swap %d0 - or.l %d1,%d0 /* data now in all 4 bytes of %d0 */ - - move.l %a1,%d1 - and.l #0xFFFFFFFC,%d1 /* %d1 = last long bound */ - cmp.l %d1,%a1 /* any bytes to set? */ - bls.b .end_b1 /* no: skip byte loop */ - - /* leading byte loop: sets 0..3 bytes */ -.loop_b1: - move.b %d0,-(%a1) /* store byte */ - cmp.l %d1,%a1 /* runs %a1 down to last long bound */ - bhi.b .loop_b1 - -.end_b1: - moveq.l #31,%d1 - add.l %a0,%d1 - and.l #0xFFFFFFF0,%d1 /* %d1 = first line bound + 16 */ - cmp.l %d1,%a1 /* at least one full line to fill? */ - blo.b .no_lines /* no, jump to longword loop */ - - mov.l %a1,%d1 - and.l #0xFFFFFFF0,%d1 /* %d1 = last line bound */ - cmp.l %d1,%a1 /* any longwords to set? */ - bls.b .end_l1 /* no: skip longword loop */ - - /* leading longword loop: sets 0..3 longwords */ -.loop_l1: - move.l %d0,-(%a1) /* store longword */ - cmp.l %d1,%a1 /* runs %a1 down to last line bound */ - bhi.b .loop_l1 - -.end_l1: - move.l %d2,-(%sp) /* free some registers */ - move.l %d3,-(%sp) - - move.l %d0,%d1 /* spread data to 4 data registers */ - move.l %d0,%d2 - move.l %d0,%d3 - lea.l (15,%a0),%a0 /* start address += 15, acct. for trl. data */ - - /* main loop: set whole lines utilising burst mode */ -.loop_line: - lea.l (-16,%a1),%a1 /* pre-decrement */ - movem.l %d0-%d3,(%a1) /* store line */ - cmp.l %a0,%a1 /* runs %a1 down to first line bound */ - bhi.b .loop_line - - lea.l (-15,%a0),%a0 /* correct start address */ - move.l (%sp)+,%d3 /* restore registers */ - move.l (%sp)+,%d2 - - move.l %a0,%d1 /* %d1 = start address ... */ - addq.l #3,%d1 /* ... +3, account for possible trailing bytes */ - cmp.l %d1,%a1 /* any longwords left */ - bhi.b .loop_l2 /* yes: jump to longword loop */ - bra.b .no_longs /* no: skip loop */ - -.no_lines: - move.l %a0,%d1 /* %d1 = start address ... */ - addq.l #3,%d1 /* ... +3, account for possible trailing bytes */ - - /* trailing longword loop */ -.loop_l2: - move.l %d0,-(%a1) /* store longword */ - cmp.l %d1,%a1 /* runs %a1 down to first long bound */ - bhi.b .loop_l2 - -.no_longs: - cmp.l %a0,%a1 /* any bytes left? */ - bls.b .end_b2 /* no: skip loop */ - - /* trailing byte loop */ -.loop_b2: - move.b %d0,-(%a1) /* store byte */ - cmp.l %a0,%a1 /* runs %a1 down to start address */ - bhi.b .loop_b2 - -.end_b2: - move.l %a0,%d0 /* return start address */ - rts - -.end: - .size memset,.end-memset diff --git a/firmware/target/coldfire/memset16-coldfire.S b/firmware/target/coldfire/memset16-coldfire.S deleted file mode 100644 index 1673038d03..0000000000 --- a/firmware/target/coldfire/memset16-coldfire.S +++ /dev/null @@ -1,146 +0,0 @@ -/*************************************************************************** - * __________ __ ___. - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ - * \/ \/ \/ \/ \/ - * $Id$ - * - * Copyright (C) 2006 by Jens Arnold - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation; either version 2 - * of the License, or (at your option) any later version. - * - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY - * KIND, either express or implied. - * - ****************************************************************************/ -#include "config.h" - - .section .icode,"ax",@progbits - - .global memset16 - .type memset16,@function - -/* Fills a memory region with specified word value - * Start address must be word aligned, length is in words - * This version is optimized for speed - * - * arguments: - * (4,%sp) - start address - * (8,%sp) - data - * (12,%sp) - length - * - * return value: - * %d0 - start address - * - * register usage: - * %d0 - data (spread to both words when using long stores) - * %d1 - temporary / data (for burst transfer) - * %d2 - data (for burst transfer) - * %d3 - data (for burst transfer) - * %a0 - start address - * %a1 - current address (runs down from end to start) - * - * For maximum speed this routine uses both long stores and burst mode, - * storing whole lines with movem.l. The routine fills memory from end - * to start in order to ease returning the start address. - */ -memset16: - move.l (4,%sp),%a0 /* start address */ - move.l (8,%sp),%d0 /* data */ - move.l (12,%sp),%a1 /* length */ - add.l %a1,%a1 - add.l %a0,%a1 /* %a1 = end address */ - - move.l %a0,%d1 - addq.l #6,%d1 - and.l #0xFFFFFFFC,%d1 /* %d1 = first long bound + 4 */ - cmp.l %d1,%a1 /* at least one aligned longword to fill? */ - blo.b .no_longs /* no, jump directly to word loop */ - - and.l #0xFFFF,%d0 /* start: spread data to both words */ - move.l %d0,%d1 - swap %d1 - or.l %d1,%d0 /* data now in both words */ - - move.l %a1,%d1 - and.l #0xFFFFFFFC,%d1 /* %d1 = last long bound */ - cmp.l %d1,%a1 /* one extra word? */ - bls.b .end_w1 /* no: skip */ - - move.w %d0,-(%a1) /* set leading word */ - -.end_w1: - moveq.l #30,%d1 - add.l %a0,%d1 - and.l #0xFFFFFFF0,%d1 /* %d1 = first line bound + 16 */ - cmp.l %d1,%a1 /* at least one full line to fill? */ - blo.b .no_lines /* no, jump to longword loop */ - - mov.l %a1,%d1 - and.l #0xFFFFFFF0,%d1 /* %d1 = last line bound */ - cmp.l %d1,%a1 /* any longwords to set? */ - bls.b .end_l1 /* no: skip longword loop */ - - /* leading longword loop: sets 0..3 longwords */ -.loop_l1: - move.l %d0,-(%a1) /* store longword */ - cmp.l %d1,%a1 /* runs %a1 down to last line bound */ - bhi.b .loop_l1 - -.end_l1: - move.l %d2,-(%sp) /* free some registers */ - move.l %d3,-(%sp) - - move.l %d0,%d1 /* spread data to 4 data registers */ - move.l %d0,%d2 - move.l %d0,%d3 - lea.l (14,%a0),%a0 /* start address += 14, acct. for trl. data */ - - /* main loop: set whole lines utilising burst mode */ -.loop_line: - lea.l (-16,%a1),%a1 /* pre-decrement */ - movem.l %d0-%d3,(%a1) /* store line */ - cmp.l %a0,%a1 /* runs %a1 down to first line bound */ - bhi.b .loop_line - - lea.l (-14,%a0),%a0 /* correct start address */ - move.l (%sp)+,%d3 /* restore registers */ - move.l (%sp)+,%d2 - - move.l %a0,%d1 /* %d1 = start address ... */ - addq.l #2,%d1 /* ... +2, account for possible trailing word */ - cmp.l %d1,%a1 /* any longwords left */ - bhi.b .loop_l2 /* yes: jump to longword loop */ - bra.b .no_longs /* no: skip loop */ - -.no_lines: - move.l %a0,%d1 /* %d1 = start address ... */ - addq.l #2,%d1 /* ... +2, account for possible trailing word */ - - /* trailing longword loop */ -.loop_l2: - move.l %d0,-(%a1) /* store longword */ - cmp.l %d1,%a1 /* runs %a1 down to first long bound */ - bhi.b .loop_l2 - -.no_longs: - cmp.l %a0,%a1 /* any words left? */ - bls.b .end_w2 /* no: skip loop */ - - /* trailing word loop */ -.loop_w2: - move.w %d0,-(%a1) /* store word */ - cmp.l %a0,%a1 /* runs %a1 down to start address */ - bhi.b .loop_w2 - -.end_w2: - move.l %a0,%d0 /* return start address */ - rts - -.end: - .size memset16,.end-memset16 diff --git a/firmware/target/coldfire/strlen-coldfire.S b/firmware/target/coldfire/strlen-coldfire.S deleted file mode 100644 index 765969da04..0000000000 --- a/firmware/target/coldfire/strlen-coldfire.S +++ /dev/null @@ -1,71 +0,0 @@ -/*************************************************************************** - * __________ __ ___. - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ - * \/ \/ \/ \/ \/ - * $Id$ - * - * Copyright (C) 2010 Nils Wallménius - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation; either version 2 - * of the License, or (at your option) any later version. - * - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY - * KIND, either express or implied. - * - ****************************************************************************/ - -/* size_t strlen(const char *str) */ - - .section .text,"ax",@progbits - .align 2 - .globl strlen - .type strlen, @function - -strlen: - move.l 4(%sp), %a0 /* %a0 = *str */ - move.l %a0, %a1 /* %a1 = start address */ - move.l %a0, %d0 - andi.l #3, %d0 /* %d0 = %a0 & 3 */ - beq.b 1f /* already aligned */ - jmp.l (-2,%pc,%d0.l*4) - tst.b (%a0)+ - beq.b .done - tst.b (%a0)+ - beq.b .done - tst.b (%a0)+ - beq.b .done - - 1: - move.l (%a0)+, %d0 /* load %d0 increment %a0 */ - /* use trick to test the whole word for null bytes */ - move.l %d0, %d1 - subi.l #0x01010101, %d1 - not.l %d0 - and.l %d1, %d0 - andi.l #0x80808080, %d0 - beq.b 1b /* if the test was false repeat */ - - /* ok, so the last word contained a 0 byte, test individual bytes */ - subq.l #4, %a0 - tst.b (%a0)+ - beq.b .done - tst.b (%a0)+ - beq.b .done - tst.b (%a0)+ - beq.b .done - /* last byte must be 0 so we don't need to load it, so we don't increment a0 - so we jump past the subq instr */ - .word 0x51fa /* trapf.w, shadow next instr */ - -.done: - subq.l #1, %a0 /* %a0 is 1 too large due to the last increment */ - sub.l %a1, %a0 /* how many times did we repeat? */ - move.l %a0, %d0 /* return value in %d0 */ - rts - .size strlen, .-strlen - |